The PHOENICS research project, in which the Fraunhofer Heinrich Hertz Institute (HHI) is a partner, kicked off this spring. The project, funded by the EU with 5.8 million euros, brings together world leaders in photonic computing to achieve energy-efficient petascale computing power with ultra-high bandwidth for the first time using light. Such processing power is essential to unlock the full potential of artificial intelligence (AI). Over the course of the four-year funding period, the consortium aims to establish so-called photonic computing as a competitive approach to machine learning.
AI applications are penetrating more and more areas of the digitized society. In doing so, the demands placed on electronic hardware in terms of computing power and storage capacity are enormous. Further development of AI requires an extensive increase in computing power (and thus transistor density); more than five times higher than that specified by Moore's Law. This law states that the transistor density on a microprocessor doubles almost every two years.
Researchers in the PHOENICS project will use innovative hardware approaches to break Moore's Law and process the enormous amounts of data needed for sophisticated AI applications. Switching from electronic to photonic approaches will enable new methods for ultra-fast information processing. To this end, photonic neuromorphic processors are the key innovation. Neuromorphic computing refers to analog computing units that learn by emulating the human brain and mimic the plasticity of a biological nervous system. For AI computing, these units promise unprecedented computing power and energy efficiency.
The acronym PHOENICS stands for "Photonic enabled petascale in-memory computing with femtojoule energy consumption" and summarizes the goals of the project: The concept of in-memory computing enables data processing similar to the human brain by eliminating the separation between computational and storage units. Photonic technology creates high-speed data transport where electronic systems currently reach their limits.
In the project, Fraunhofer HHI is developing the part of the neuromorphic processor responsible for encoding the data. For this purpose, the team will use Fraunhofer HHI's hybrid photonic integration platform PolyBoard to separate the individual wavelengths of the signal (up to 24). Subsequently, the individual wavelengths are modulated and amplified in many parallel InP chips to encode the data, after which the signals are recombined with a second PolyBoard chip. This parallel processing of different light wavelengths achieves a very high data throughput.
The project, which is scheduled to run for four years, is a research venture coordinated by Westfälische Wilhelms-Universität Münster (Germany). Besides Fraunhofer HHI, other partners include the University of Exeter (UK), École polytechnique fédérale de Lausanne (Switzerland), Nanoscribe GmbH & Co. KG (Germany), the University of Oxford (UK), the University of Ghent (Belgium), IBM Research GmbH (Switzerland) and MicroR Systems.